Future directions
Experiments and discussion
Experiments and discussion
Markov chain models are built up for Mozart's and Haydn's scale degree class
transitions. The repertoire consists of all of their string quartets, 100
movements from Mozart and 212 movements from Haydn.
The scale degree is defined relative to the tonics, i.e., all works are
transposed into the same key. Also, for the sake of reducing the
dimensionality of transition matrice, the system does not distinguish between
the same degree class at different octaves. The system does not
distinguish between major keys and minor keys, either.
Then, the two-way identification
tests based on Kullback-Leibler distances were conducted for each of the
4 voices. The tests were conducted in a bootstrap manner. In other words,
it is always assumed that the computer has been exposed to the
all the works in the repertoire except the one to be identified.
Table 1:
Mozart vs. Haydn identification tests.
Part |
Mozart |
Haydn |
ViolinI |
68.0% |
64.2% |
ViolinII |
58.0% |
64.2% |
Viola |
61.0% |
53.8% |
Cello |
57.0% |
52.8% |
|
Table 1 shows the result of two-way identification tests.
The boldfont indicates data that are statistically significant.
To qualify for statistical significance, a recognition rate has to be
higher than
of random flips of fair coins, where
is the mean, 50%; and is the standard deviation,
which is inversely proportional to the square root of the number of flips,
according to the central limit theorem.
As a controlled experiment, repertoires are defined as two
randomly assigned, and mutually exclusive lists of quartets.
Then, the two-way style identification tests are conducted, and the
result is shown in Table 2. Each of the data shows an average
of ten runs.
Table 2:
Random repertoires identification tests.
Part |
Random A |
Random B |
ViolinI |
42.6% |
56.1% |
ViolinII |
44.5% |
53.4% |
Viola |
47.4% |
51.9% |
Cello |
43.1% |
53.4% |
|
Human listening tests are conducted via web survey. A user is played,
the MIDI piano version of, a
quartet from either composer by equal chance, and is asked to identify
the composer. As of June 11, 2002, web users average 59.0%
of accuracy in 1865 attempts.
How experienced the users are with the repertoire
ranges from novices of classical music to string players that have
played some of the works.
Future directions
Experiments and discussion
Experiments and discussion
Copyright © 2002-06-11
Center for Computer Research in Music and Acoustics,
Stanford University